Your mouth needs training your eyes cannot give it. Italian asks your articulators to do things Dutch and English never ask: the rolled r, the double consonants that separate pena from penna, seven vowels held pure to the end of the word. You can know all of this and still be unable to do it at speed, for the same reason reading about the piano does not play it. The two cheapest tools for the physical side of speaking are reading aloud and shadowing, and both work alone in a room.
Reading aloud
Take a text at or below your level (a page of a graded reader you have already read silently is ideal) and read it aloud, at natural speed, for five minutes. That is the whole method. What it trains: the mapping from Italian spelling to Italian sound (mercifully regular, one of the language’s gifts), the motor patterns of connected speech, and the habit of finishing your vowels. What it does not train: choosing your own words; for that, see the output hypothesis. It is pronunciation and fluency work wearing a speaking costume, and it is worth doing precisely because the words are provided: with word-choice removed, all your attention goes to sound.
Two upgrades. Read the same page on consecutive days and mark it against the clock; now it is also fluency practice. And record yourself once a week; the gap between how you think you sound and the recording is uncomfortable and extremely instructive.
Shadowing
Shadowing is the intense cousin: play native audio and speak along with it, echoing a half-second behind, matching rhythm and intonation as you go. Research interest in the technique has grown steadily: Hamada’s studies with Japanese learners of English report gains in listening and in the fluency of repetition, and the mechanism is plausible: shadowing forces you to process sound in real time and reproduce it before your inner translator can interfere.
Make it survivable: use short audio (30–60 seconds) you already understand: a dialogue you have studied, a paragraph of an audiobook you have read. Listen once, shadow three or four times, then once with the transcript to catch what you were mangling. Italian is generous material for this: its syllable-timed rhythm and open vowels make the melody easier to catch than French or English. Do not shadow the news at C2 speed at A2; frustration is not a method.
What speech recognition can and cannot check
In 
amo Italian, the Ad alta voce feature runs this loop with feedback: you read a passage aloud and on-device Italian speech recognition (free, offline, nothing sent to a server) verifies whether it understood the words you produced. That is a genuinely useful signal: if the recogniser heard pena when you meant penna, your double consonant needs work.
And the honest limit, stated plainly: recognising words is not judging accent. Fine-grained feedback on intonation and stress placement is beyond what I am willing to promise from on-device tools, and I would rather tell you that than sell you a pronunciation score I cannot stand behind. For the fine grain, record yourself, compare against the native audio, and occasionally borrow the ears of a native: a teacher’s ear in a conversation lesson catches in seconds what software misses for months.
References and attributions: Hamada, “Shadowing: Who benefits and how?”, Language Teaching Research, 2016. Findings are paraphrased from the published literature with attribution. Part of the Science of Learning Italian series.
